Calculate best thresholds

calculate_thresholds.py

import numpy as np
from sklearn.metrics import roc_curve

def calculate_thresholds(n_classes, y_true, y_pred):
    """
    n_classes => 123
    y_true => [[1, 0], ...] # one hot encode lavels
    y_pred => np.array([[0.9, 0.3]]) # with probabilities
    """
    fpr = dict()
    tpr = dict()
    thresholds = dict()
    best_thresholds = dict()
    
    for i in n_classes:
        fpr[i], tpr[i], thresholds[i] = roc_curve(y_true[:, i], y_pred[:, i])
        tf = tpr[i]-(1-fpr[i])
        best_thresholds[i] = 1-thresholds[i][np.absolute(tf).argsort()[0]]
        
    return {
        'fpr': fpr,
        'tpr': tpr,
        'thresholds': thresholds,
        'best_thresholds': best_thresholds
    }